PRAGMADiscover your offer
Articles

Zero-retention AI providers: what their official texts state

Painting: a keeper raises an antique hourglass whose sand turns to gold dust and fades away

What does Zero Data Retention mean for an AI provider?

Zero Data Retention (ZDR) processes prompts and responses in volatile memory without durable writes after returning. By default, commercial APIs exclude training yet keep 30- to 90-day abuse logs. ZDR requires prior approval, stateless endpoints, and checking model exceptions.

Key points

  • Disabling training differs from Zero Data Retention1: by default, several APIs retain abuse logs (up to 30 days at OpenAI2).
  • ZDR covers stateless calls only: files, batches and retention-bound models stay excluded at Anthropic3 or blocked on Bedrock4.
  • To stay in Europe, one must choose, when the provider offers one, a regional endpoint excluding Global; transfers to the United States (outside Order No. 2026-5485) require an Article 98 safeguard such as the APDP Standard Protection Clauses6, or Article 99 or 100.
  • Law and safety override ZDR (up to 2 years per flagged session at Anthropic3), requiring a contract (Law No. 1.565, Article 267) and audits (Ordinance No. 11.327, Article 88).

Linking a model to email or a CRM in Monaco engages the full data lifecycle. Without ranking or recommending any brand, this page compiles dated official texts from OpenAI, Anthropic, Google Cloud, Microsoft, Mistral AI and AWS on ZDR, Europe and abuse logs, alongside data sovereignty and AI in Monaco.

Zero retention and non-training: two controls

Opting out of training protects weights; Zero Data Retention (ZDR) bars storage at rest after the call. Mistral AI (18 November 2026) states that ZDR and training opt-out are separate1, and OpenAI distinguishes default API non-training since 1 March 2023 from 30-day abuse logs2.

Under Law No. 1.565 of 3 December 20247 (AI and personal data in Monaco), Article 4 limits retention, Article 27 logs retention periods in the register, and Article 26 requires a written processor contract with final deletion (like Article 28 of the GDPR9); Article 8 of Sovereign Ordinance No. 11.327 of 10 July 20258 obliges the processor to provide the information demonstrating compliance with Article 26 and to allow audits.

Direct providers: OpenAI, Anthropic and Mistral AI

OpenAI (20 November 2026) states the API never trains models without explicit opt-in2, yet keeps 30-day abuse logs and `/v1/responses` state if `store=true`. Once approved, ZDR removes abuse logs and forces `store=false` on `/v1/chat/completions` and `/v1/responses`, excluding stateful calls (`/v1/files`, `/v1/batches`, `/v1/vector_stores`, `/v1/assistants`, `/v1/conversations`).

At Anthropic, Section B of the Commercial Terms (17 June 2025)10 bars commercial training (Anthropic Ireland, Limited: EEA, Switzerland, UK). Per the API retention guide (2026)3 and 15 October 2026 note11, ZDR stops storage after `/v1/messages`, excluding Console, `/v1/files`, `/v1/messages/batches` (29 days), code (30 days) and Covered Models (`Claude Fable 5` to `5.2`, `Claude Mythos 5` to `5.2`, 30 days unless authorized).

Mistral AI (18 November 2026) notes that Vibe outside Enterprise trains by default until opted out12 and Studio and API toggles turn off separately (excluding Labs, Public Preview, commented feedback). Mistral Studio ZDR1, on approved pay-as-you-go accounts, covers stateless endpoints (`/v1/chat/completions`, `/v1/fim/completions`, `/v1/embeddings`, `/v1/moderations`, `/v1/classifications`, `/v1/ocr`, audio), excluding `/v1/files`, batches, agents and Vibe.

Clouds: Google Vertex AI, Microsoft Foundry and AWS Bedrock

Google Cloud documents Vertex AI under Gemini Enterprise Agent Platform (20 November 2026): no training without instruction13. ZDR requires disabling 24-hour caching (`disableCache: true` or `EPHEMERAL`), leaving BigQuery off, setting `store = false` (Interactions API), and avoiding Grounding Search (3 days) and Maps (30 days); a Cloud Master Agreement exempts accounts by default from 90-day logs14.

On Azure OpenAI (Foundry Models sold by Azure, 17 November 2026), Microsoft confirms prompts and outputs stay unavailable to OpenAI and out of training15. Modified Abuse Monitoring16 disables storage and review (`ContentLogging: false`): retained before 1 November 2026, it ends for models released from 1 November 2026 after 30 days' notice.

On Amazon Bedrock, AWS confirms (privacy, 18 November 202617; protection, 20 November 202618) zero provider sharing and zero base-model training. The retention guide (20 November 2026)4 provides regional mode `none` (IAM or SCP), blocking durable writes, `store=true` and retention-bound models; S3 or CloudWatch logging19 is off by default.

EU processing and transfers from Monaco

From Monaco, location governs transfers (AI hosting in Europe and Monaco, Monaco Cloud and sovereign cloud). Under Article 97 of Law No. 1.5657 and Order No. 2026-548 of 30 September 20265, the EU, EEA, Switzerland and the UK are adequate, not the US: Article 98 requires safeguards such as the APDP Standard Protection Clauses of 11 November 2026 (Deliberation No. 2026-24)6.

Mistral AI (18 November 2026) states data is hosted by default in the EEA20, excluding the US endpoint or subprocessors under GDPR clauses. OpenAI (20 November 2026) offers Europe (EEA, Switzerland) residency via `eu.api.openai.com`21 with Cloudflare TLS (10% uplift for models released on or after 5 March 2026), excluding system data. Anthropic limits API residency (2026)22 to `us` or `global`, pointing to Google Cloud or Bedrock in Europe.

Across clouds, Google Cloud residency (20 November 2026)23 separates `eu` (EU only) from `global`; Microsoft processes regional deployments in an EU Data Boundary (22 July 2026)24 region within that boundary; an EU DataZone deployment may be processed in any EU Member State; Global and Batch are excluded15 (EEA reviewers for EEA deployments); AWS Bedrock contrasts (cross-Region, 20 November 202625) `eu.` with global profiles.

Safety exceptions and residual traces

ZDR yields to law and filters: Anthropic (API retention3) may keep flagged sessions up to 2 years and keeps User Safety scores11; OpenAI (controls2) retains CSAM alerts and may notify safety retention. On Bedrock26, AWS retains `Claude Fable 5` to `5.2` up to 30 days (outside EFS through at least 30 June 2027) and classifier-flagged OpenAI traffic (excluding `gpt-oss`) up to 30 days; Google Cloud logs Frontier Models 30 days in customer environments14.

Without dedicated settings, 24-hour caches (OpenAI2, Google Cloud13), 24-hour JSON schemas at Anthropic3 or outside residency at OpenAI21, billing and customer logs persist (Article 31, Law No. 1.5657).

Six questions for a provider and PRAGMA's framework

Before opening files to an AI API, obtain six written proofs.

At PRAGMA, every outgoing message is validated by a person, each agent's access is scoped, every action is logged, and client data never trains an external model. Your tools run today on European servers that we administer ourselves. Hosting in Monaco is available on request, subject to technical and contractual validation. We check these points during an AI audit in Monaco, notably for family offices and firms.

This article provides general information from the official texts cited and is not legal advice: for your case, consult a lawyer or the APDP.

Frequently asked questions

Does disabling training equal Zero Data Retention?

No: it protects weights, while 30- to 90-day logs persist without ZDR or Modified Abuse Monitoring.

Is store=false enough for ZDR?

No: per AWS Bedrock4 and Google Cloud13, `store=false` stops application history, not abuse logs.

Does a Global endpoint stay in the EU?

No: on Google Cloud, Microsoft Foundry and AWS Bedrock, `Global` can leave Europe; require an EU regional endpoint.

What can a vendor keep despite ZDR?

Legal obligations, sessions flagged by safety systems (up to 2 years at Anthropic3), CSAM alerts at OpenAI2, billing data, and stateful features called.

Why do US transfers need specific clauses?

Absent from Order No. 2026-5485, the US falls under Article 98 of Law No. 1.565 and APDP Clauses.

To audit your AI data flows under these rules, describe your tools via the contact form.

Request the free audit and prototype

Sources

  1. Mistral AI, Can I activate Zero Data Retention (ZDR)?,
  2. OpenAI, Data controls in the OpenAI platform,
  3. Anthropic, API and data retention | Claude Platform Docs,
  4. AWS, Data retention | Amazon Bedrock User Guide,
  5. Legimonaco, Arrêté ministériel n° 2026-548 du 30 septembre 2026 fixant la liste des pays, territoires et organisations internationales disposant du niveau de protection adéquat,
  6. APDP, Transferts de données : Clauses Types de Protection,
  7. Legimonaco, Loi n° 1.565 du 3 décembre 2024 relative à la protection des données personnelles,
  8. Legimonaco, Ordonnance Souveraine n° 11.327 du 10 juillet 2025 portant application de la loi n° 1.565 du 3 décembre 2024,
  9. EUR-Lex, Règlement (UE) 2016/679 du Parlement européen et du Conseil du 27 avril 2016 (RGPD),
  10. Anthropic, Commercial Terms of Service,
  11. Anthropic, I have a zero data retention agreement with Anthropic. What products does it apply to?,
  12. Mistral AI, Do you use my user data to train your Artificial Intelligence models?,
  13. Google Cloud, Gemini Enterprise Agent Platform and zero data retention,
  14. Google Cloud, Abuse monitoring | Gemini Enterprise Agent Platform,
  15. Microsoft, Data, privacy, and security for Models sold by Azure in Microsoft Foundry,
  16. Microsoft, Foundry Models sold by Azure abuse monitoring,
  17. AWS, Security, privacy, and responsible AI in Amazon Bedrock,
  18. AWS, Data protection | Amazon Bedrock User Guide,
  19. AWS, Monitor model invocation using CloudWatch Logs and Amazon S3,
  20. Mistral AI, Where do you store my data or my Organization's data?,
  21. OpenAI, Data residency | OpenAI API,
  22. Anthropic, Data residency | Claude Platform Docs,
  23. Google Cloud, Data residency | Gemini Enterprise Agent Platform,
  24. Microsoft, What is the EU Data Boundary?,
  25. AWS, Route model inference requests across AWS Regions with cross-Region inference,
  26. AWS, Amazon Bedrock abuse detection,

Add PRAGMA as a preferred source on Google